BGE-Large-En-V1.5-ONNX-O4

This is an ONNX O4 strategy optimized version of BAAI/bge-large-en-v1.5 optimal for Cuda. It should be much faster than the original version.

Usage

# pip install "optimum[onnxruntime-gpu]" transformers

from optimum.onnxruntime import ORTModelForFeatureExtraction
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('hooman650/bge-large-en-v1.5-onnx-o4')
model = ORTModelForFeatureExtraction.from_pretrained('hooman650/bge-large-en-v1.5-onnx-o4')
model.to("cuda")

pairs = ["pandas usually live in the jungles"]
with torch.no_grad():
    inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)
    sentence_embeddings = model(**inputs)[0][:, 0]

# normalize embeddings
sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, dim=1)